Skip to content
Text Dalia
Nutrition

How accurate are photo calorie estimates?

What studies of AI photo estimates actually found, why big plates and hidden food trip them up, and why the error that matters most is the one that always points the same way.

EVEashan VagishOctober 15, 20267 min read

Last reviewed October 15, 2026

Plates of grilled meat, greens and sauces seen from above on a wooden table.

Photo calorie estimates are rough. In a 2025 test on 52 food photos, the two best AI models missed calories by 35.8% on average, and all of them read large portions low.[1] Self-reported food logs miss too: most studies that check them against a lab measure of what people burn find under-reporting.[2]

Key takeaways

  • In a 2025 test on 52 standard food photos, the two best AI models missed calories by 35.8% on average.[1]
  • Every model in that test read big portions too low, and the gap grew with portion size.[1]
  • Self-reported intake is not error-free: most studies that check it against a lab measure find people report eating less than they do.[2]
  • A photo method checked against a lab measure over 6 days came within 152 calories a day of it on average, with wide swings from person to person.[16]
  • Self-monitoring goes with more weight loss in studies, though the evidence has limits.[18, 19]

What the studies found

Researchers test photo estimates on meals whose contents they already know, from weighing, recipes or survey data. Then they compare the estimate with that number. The studies below use different foods, models and measures, so their numbers are not directly comparable.

StudyWhat was testedWhat it found
Sweden, 20253 AI models on 52 standard photos, checked against weighed foodChatGPT and Claude missed calories by 35.8% on average. Gemini missed by 64.2% to 109.9% across nutrients.[1]
Ireland, 2025ChatGPT-4 on 114 meal photos from a national diet surveyNamed the foods correctly 93.0% of the time. Portion weights were good for small meals, poor for medium and large ones.[3]
Japan, 202610 AI models and 10 dietitians on 15 hospital meals, photographed under set light and angleFor calories, the top models and the dietitians were often within 10%. All the models overestimated fat by more than 20%.[4]
Taiwan, 2025714 photos from 3-day weighed food records of 57 young adultsA custom GPT-4o tool estimated energy within 10% to 20%. Common errors were portion sizes and food hidden from view.[5]
Peru, 20263 sizes of the open Gemma-3 model on 510 lunch photos and meal names from a cookbookThe best model was off by 108 calories per serving on average. The authors said that is not yet good enough for clinical or consumer use.[6]
Recent tests of AI calorie estimates from food photos

In the Irish test, ChatGPT-4 named foods well and ranked meals by nutrient content.[3] In the Swedish test, the error grew with the size of the plate.[1]

Why a photo misses

A photo has to stand in for a scale. In the studies, portion size and food out of view were common sources of error.[5, 1]

  • Big portions read low. All three models in the Swedish test underestimated more as portions grew.[1] ChatGPT-4 estimated meal weight well for small meals and poorly for medium and large ones.[3]
  • Fat is hard to see. In the Japanese test, every AI model overestimated fat by more than 20%, which the authors called an "invisible nutrient" bias.[4]
  • Food out of view is missed. In the Taiwanese study, food hidden in the photo and wrong portion sizes were common errors.[5]
  • Photos taken late miss food. A review found photo methods can be prone to underestimation when people do not take a clear photo before they finish eating.[7]

Context helps. When ChatGPT-5 got a photo plus a list of ingredients with amounts, its calorie estimates improved.[8] With the same ingredient list and no photo, they got worse again, so the image itself carried useful information.[8]

How hand logging compares

Self-reported food records and recalls, the older ways to track, have their own errors. A review of 59 studies in 6,298 adults found that most reported significant under-reporting of calories when checked against a lab measure of energy burned.[2]

In a US study of 484 adults aged 40 to 69, men under-reported calories by 12% to 14% on 24-hour recalls.[9] Women under-reported by 16% to 20%.[9] In a study of 10 people with obesity who said they could not lose weight, reported food intake was 47% below their actual intake on average.[10]

The numbers you log from are not exact either. US rules treat a labeled food as misbranded when its calories come in more than 20% above the label.[11] In 269 restaurant foods, 19% had at least 100 calories more than stated.[12] A test of 24 snack foods found calories 4.3% above the label, which was still well within the limits.[13]

One study of 50 people in Belgium used a cutoff to remove extremely high, likely wrong MyFitnessPal values, 2.8% of them.[14] After that, its calorie totals tracked a national food database closely (r=0.96).[14]

Note

Photos versus traditional methods

A 2020 meta-analysis of 13 studies found that image-based methods under-reported calories by 179.32 calories on average.[15] Against a lab measure of energy burned, the gap was 448.04 calories.[15] Against traditional food records and recalls, there was no statistical difference. The authors concluded that photo methods, like traditional ones, have serious measurement errors.[15]

One meal versus a whole week

In a study of the Remote Food Photography Method, 50 adults photographed their meals for 6 days.[16] Their intake came within 152 calories a day of a lab measure of energy burned on average, a gap that was not statistically significant.[16]

The spread between people was wide, though: the standard deviation was 694 calories a day.[16] Reminders also mattered. In an earlier group who got fewer reminders to take photos, the estimate fell 895 calories a day short, against 270 for those reminded more.[16]

Worked example

When errors cancel, and when they add up

Take someone eating 1,800 calories a day in three meals of 600. Suppose their photo estimates are off by up to 150 calories a meal. Two cases show why the direction of the error matters more than its size.

  1. Errors go both ways

    (600 + 150) + (600 − 150) + 600

    1,800 logged

  2. Gap from the real day

    1,800 − 1,800

    0 cal

  3. Every meal reads low

    (600 − 150) × 3

    1,350 logged

  4. Gap from the real day

    1,800 − 1,350

    450 cal a day

  5. Gap over a week

    450 × 7

    3,150 cal

What the log misses in the second case

450 calories a day

In this example, a person planning a 500-calorie daily cut would really be cutting only 50. The National Academies describe setting energy intake as a two-step process: estimate, then watch body weight over time.[17] If weight gain or loss does not go as planned, they say to adjust intake as needed.[17]

Does tracking help if the numbers are rough?

A review of 22 weight loss studies found a consistent link between self-monitoring and weight loss.[18] The authors rated the evidence as weak, mostly because the samples were narrow and the measures were self-reported.[18]

In a 24-week online program, people who lost at least 5% of their weight opened their food log more often: 2.4 times a day versus 1.6.[19] The time spent logging fell from 23.2 minutes a day in month 1 to 14.6 minutes in month 6.[19] These are associations, so they do not prove that logging caused the loss.

How to get a better estimate from a photo

  1. 1Take the photo before you eat. Photo methods can underestimate when the image comes after food is gone.[7]
  2. 2Get the whole plate in view. Hidden food was a common source of error.[5]
  3. 3Add what the camera cannot see. Name the oil, butter, dressing or sauce, and the amount if you know it. Ingredient details improved estimates.[8]
  4. 4Watch the big plates. Large portions are where estimates read lowest.[1]
  5. 5Set reminders. In one study, people who got fewer reminders to take photos came up much further short.[16]

For the targets these estimates feed into, see calories and macros, explained and how to count macros. For a side-by-side of tracking apps, see the best calorie tracker apps.

When to talk to a clinician

Frequently asked questions

In a 2025 test, the best AI models missed meal calories by 35.8% on average, and large portions read low. In a 6-day study, one photo method came within 152 calories a day of a lab measure on average, though people varied widely. Compare what your log predicts with your weight over several weeks, and adjust if they disagree.

Not directly. In the studies, portion size and food hidden from view were common sources of error. In one study, adding a list of ingredients and amounts to the photo improved the estimates.

Small errors that go both ways matter little over a week. A steady error in one direction matters more: underestimating every meal by 150 calories hides 450 calories a day. Under US rules, a labeled food counts as misbranded only once its calories run more than 20% over the label, so labels are not exact either.

References & further reading

  1. 1.Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images, Current Developments in Nutrition (Fridolfsson et al.), 2025
  2. 2.Validity of Dietary Assessment Methods When Compared to the Method of Doubly Labeled Water: A Systematic Review in Adults, Frontiers in Endocrinology (Burrows et al.), 2019
  3. 3.An Evaluation of ChatGPT for Nutrient Content Estimation from Meal Photographs, Nutrients (O'Hara et al.), 2025
  4. 4.Accuracy of AI-Based Nutrient Estimation from Standardized Hospital Meal Images: A Comparison with Registered Dietitians, Nutrients (Isobe et al.), 2026
  5. 5.Customized multimodal Diabot-GPT-4o enhances accuracy of image-based dietary assessments in dietetic trainees in Taiwan: validation against weighed food records, American Journal of Clinical Nutrition (Chen et al.), 2025
  6. 6.Evaluating Large Language Model Accuracy in Predicting Peruvian Meal Nutrition, Journal of Nutrition (Carrillo-Larco et al.), 2026
  7. 7.Image-assisted dietary assessment: a systematic review of the evidence, Journal of the Academy of Nutrition and Dietetics (Gemming et al.), 2015
  8. 8.Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios, Nutrients (Rodríguez-Jiménez et al.), 2025
  9. 9.Using intake biomarkers to evaluate the extent of dietary misreporting in a large sample of adults: the OPEN study, American Journal of Epidemiology (Subar et al.), 2003
  10. 10.Discrepancy between self-reported and actual caloric intake and exercise in obese subjects, New England Journal of Medicine (Lichtman et al.), 1992
  11. 11.21 CFR 101.9, Nutrition labeling of food, US Code of Federal Regulations (Legal Information Institute)
  12. 12.Accuracy of stated energy contents of restaurant foods, JAMA (Urban et al.), 2011
  13. 13.Food label accuracy of common snack foods, Obesity (Jumpertz et al.), 2013
  14. 14.Accuracy of Nutrient Calculations Using the Consumer-Focused Online App MyFitnessPal: Validation Study, Journal of Medical Internet Research (Evenepoel et al.), 2020
  15. 15.Validity of image-based dietary assessment methods: A systematic review and meta-analysis, Clinical Nutrition (Ho et al.), 2020
  16. 16.Validity of the Remote Food Photography Method (RFPM) for estimating energy and nutrient intake in near real-time, Obesity (Martin et al.), 2012
  17. 17.Dietary Reference Intakes for Energy, Summary, National Academies of Sciences, Engineering, and Medicine, 2023
  18. 18.Self-monitoring in weight loss: a systematic review of the literature, Journal of the American Dietetic Association (Burke et al.), 2011
  19. 19.Log Often, Lose More: Electronic Dietary Self-Monitoring for Weight Loss, Obesity (Harvey et al.), 2019
  20. 20.Carbohydrate Estimation Accuracy of Two Commercially Available Smartphone Applications vs Estimation by Individuals With Type 1 Diabetes: A Comparative Study, Journal of Diabetes Science and Technology (Baumgartner et al.), 2024
  21. 21.Calorie counting and fitness tracking technology: Associations with eating disorder symptomatology, Eating Behaviors (Simpson and Mazzeo), 2017
  22. 22.Eating Disorders, National Institute of Mental Health
EV

Eashan Vagish

Founder, Dalia Health

Eashan Vagish is the founder of Dalia Health, where he works on making heart and metabolic health easier to understand and track. He writes these guides to answer the health questions people actually ask, in plain language, with links to the research behind every claim.

About this guide: These guides are written by the Dalia Health team. Every factual claim links to its source, preferring peer-reviewed research and public-health bodies such as the NIH, CDC and WHO. Each guide shows when it was last reviewed. They are for general education and are not a substitute for medical advice. Read our editorial policy at daliahealth.com/editorial-policy, and report an error to info@daliahealth.com.

Medical disclaimer: This article is for general educational purposes only and is not medical advice. It is not a substitute for diagnosis or treatment from a qualified healthcare professional. Always talk to your doctor about your individual health, and seek immediate care for any urgent symptoms.

How we write, source and correct these guides: our editorial policy. Spot an error? Email info@daliahealth.com.

Keep reading

How to count macros, step by step

How macro targets are set, how calories turn into grams, and how a starting split compares with the national ranges.

Nutrition 4 min read